Wayve says its safety envelope mitigates most critical on-road interventions

Wayve says its AI driving system can make its risk judgments visible—and that the mechanism mitigates the vast majority of safety-critical interventions on the road. In a November 4 engineering article, the company presented initial findings for its Learned Safety Envelope, alongside the architecture and testing methods behind its safety case.

The article, written by Simone Fabris, Wayve’s vice president of product and delivery, gives no numerical result for the intervention claim. Its technical explanation centers on a distinction: learning how to drive and learning what is safe are separate responsibilities, even when they use the same understanding of the road.

Risk that engineers can inspect

The Learned Safety Envelope is a live representation of which actions the AI Driver considers safe or unsafe, and how much safety margin remains. Wayve says it learns those judgments directly from data, rather than relying on hand-labeled risks or fixed thresholds. Engineers can compare the displayed judgment with what the vehicle actually does.

One example shows a cyclist emerging from behind a parked vehicle. According to Wayve, a red risk zone starts forming after the cyclist’s head becomes visible through the parked car’s windows for a few frames—before the cyclist and bicycle are fully visible. The example illustrates the company’s approach: respond to partial signs of danger instead of waiting to recognize a complete object and apply a predefined rule.

This does not, by itself, prove that the complete driving system is safe.

Simone Fabris, Wayve vice president of product and delivery

Three jobs, one view of the road

Wayve separates driving and risk anticipation into tasks, not separate neural networks. The driving task selects smooth, comfortable, traffic-rule-compliant movement. Risk anticipation assesses the room, time and braking distance a situation demands. Both learn end-to-end, but have distinct objectives, evaluation processes and deployment responsibilities.

The company argues that these different perspectives provide complementary checks: one focuses on behavior and intent, the other on geometry and physical limits. Wayve says that diversity makes them less vulnerable to the same underlying errors. Keeping the tasks from interfering with each other on automotive chips is another part of its safety design.

A third task, which Wayve calls self-awareness, monitors whether that shared scene understanding remains trustworthy. Unfamiliar conditions, degraded sensing or low confidence can trigger corrective action. The required response depends on the automation level: return control conservatively to a responsible human driver, maintain control during a requested handover, or reach a safe state without relying on a human fallback.

The whole vehicle still has to pass

Wayve says road mileage alone cannot efficiently establish confidence in rare, unusual events, because those situations may barely appear even across millions of miles. It also emphasizes that evaluation must cover the complete driving system, rather than treating a model’s isolated score as sufficient.

Its validation approach combines closed-loop simulation, where driving decisions affect how a simulated situation unfolds, with targeted road tests and conventional software validation. Wayve assigns each a different job: simulation supplies controlled breadth, road testing adds depth, and software checks assess whether the code and vehicle integration work as intended. Evidence is assessed against explicit release criteria to judge the risk that remains.

The company places that work alongside automotive safety standards and frameworks, including ISO 26262, ISO 21448, ISO/PAS 8800 and UL 4600. Where standards do not fully prescribe testing for an end-to-end driving model, Wayve says it adds behavioral, scenario and uncertainty testing, plus continuous monitoring. Field data informs later model releases, each of which is revalidated before deployment.